Introduction

Liberbank needed to keep its branches open under COVID-19 capacity limits, which meant enforcing appointment-only visits at a moment when its contact center wasn’t sized for the exponential spike in calls a lockdown would bring. An automated booking system had to absorb that surge instead of adding to it, and it had to ship fast: branches couldn’t operate under uncertain capacity rules for long.

That reframed what the assistant was for. Liby wasn’t built to answer questions. It was built to protect the contact center, and every design decision was measured against whether it kept a conversation from turning into a phone call.

I led conversational design, from Liby’s first booking flow through the personality and tone work that shaped it, and contributed to the product side of the rollout: a modular UI kit for design-engineering handoff, and a training program so Liberbank’s own team could maintain and extend the assistant after launch.

Framing

The first version had to solve one thing: appointment booking, the single most urgent case. We discarded launching broad FAQ coverage alongside it. Agile delivery on a scalable architecture meant Liby could ship that first flow in under two months, then layer in new use cases, like answering questions about digital banking products, once the urgent case was live and stable.

That sequencing was the core decision: solve the case with the clearest cost of failure first, and build the architecture so the next use case was an addition, not a rebuild.

Crafting

Booking flow and conversational design

Liby walked a customer straight from intent to a confirmed slot: pick the nearest office, see real availability, and book it in the same exchange, no branch call required. Every step was designed to work for someone under pandemic stress, not just an ideal user with time to explore menus.

A two-stage exit to a human

An assistant built to protect a contact center has to be deliberate about when it sends someone back to it. Handing over too early defeated the purpose; handing over too late left a customer stuck during a lockdown with no branch to walk into.

Failure was designed in two stages. The first fallback stayed inside the chat and worked to steer the conversation back onto a path Liby could complete. Only when that didn’t land did the second fallback offer a phone call with an agent. The first stage absorbed the misunderstandings that were really phrasing problems; the second reserved the contact center for the cases that genuinely needed a person.

Personality and tone

Personality and tone came out of a structured exercise across three levels, channel, geography, and cultural register, before a single conversation flow was written. That work became the style guide the rest of the team built from, so every new flow sounded like the same assistant.

Handoff and team enablement

A modular UI kit gave engineering a consistent set of components to build from instead of reinterpreting each flow’s design. Alongside it, I designed and delivered a training program so Liberbank’s own product team could maintain and extend Liby independently once the engagement ended.

Outcome

In its first six months, Liby handled 355k messages and processed 2% of all the appointment requests Liberbank received that year, without adding load to the contact center it was built to protect. 4.5% of messages ended in an error or an escalation to a human agent, and 90% of users rated their interactions positively.

The first version shipped in under two months from a standing start during a lockdown. That speed, and the scalable architecture behind it, is what let Liberbank keep adding use cases after the initial rollout instead of starting over each time.

Narrowing the scope had a cost. Intents tied to promotional topics were deliberately left out of the first version to stop them branching against the booking flow, which protected the launch but left Liby visibly incomplete on subjects the bank was actively communicating elsewhere. Those intents were added afterwards. The trade-off was the right one under the deadline, and it is still the decision I would defend: an assistant that does one thing reliably buys the credibility to do the second.